The Presidential Puzzle: Political Cycles and the Stock Market
←
→
Page content transcription
If your browser does not render page correctly, please read the page content below
THE JOURNAL OF FINANCE VOL. LVIII, NO. 5 OCTOBER 2003
The Presidential Puzzle: Political Cycles and the
Stock Market
PEDRO SANTA-CLARA and ROSSEN VALKANOV n
ABSTRACT
The excess return in the stock market is higher under Democratic than Repub-
lican presidencies: 9 percent for the value-weighted and 16 percent for the
equal-weighted portfolio. The di¡erence comes from higher real stock returns
and lower real interest rates, is statistically signi¢cant, and is robust in sub-
samples. The di¡erence in returns is not explained by business-cycle variables
related to expected returns, and is not concentrated around election dates.
There is no di¡erence in the riskiness of the stock market across presidencies
that could justify a risk premium. The di¡erence in returns through the
political cycle is therefore a puzzle.
IN THE RUN-UP TOALL PRESIDENTIAL ELECTIONS, the popular press is awash with reports
about whether Republicans or Democrats are better for the stock market. Unfor-
tunately, the popular interest has not been matched by academic research. This
paper ¢lls that gap by conducting a careful empirical analysis of the relation
between presidential elections and the stock market.
Using data since 1927, we ¢nd that the average excess return of the value-
weighted CRSP index over the three-month Treasury bill rate has been about 2
percent under Republican and 11 percent under Democratic presidentsFa strik-
ing di¡erence of 9 percent per year! This di¡erence is economically and statisti-
cally signi¢cant. A decomposition of excess returns reveals that the di¡erence is
due to real market returns being higher under Democrats by more than 5 percent,
as well as to real interest rates being almost 4 percent lower under Democrats.
The results are even more impressive for the equal-weighted portfolio, where the
di¡erence in excess returns between Republicans and Democrats reaches 16 per-
cent. Moreover, we observe an absolute monotonicity in the di¡erence between
size-decile portfolios under the two political regimes: From 7 percent for the lar-
gest ¢rms to about 22 percent for the smallest ¢rms.
n
Santa-Clara and Valkanov are from The Anderson School, University of California, Los
Angeles. We thank Antonio Bernardo, Michael Brandt, Michael Brennan, Bhagwan
Chowdhry, Brad Cornell, Eugene Fama, Shingo Goto, Mark Grinblatt, Harrison Hong, Jun
Liu, Francis Longsta¡, Monika Piazzesi, Richard Roll, and Jose¤ Tavares for useful comments.
We are especially grateful to Maria Gonzalez, Richard Green (the editor), and an anonymous
referee for many suggestions that have greatly improved this paper. We thank Kenneth French
and G. William Schwert for providing ¢nancial data. All remaining errors are our own.
18411842 The Journal of Finance
When faced with a result such as this, we have to ask ourselves whether the
¢ndings are spurious.We conduct several robustness checks, including studying
di¡erent subsamples, correcting the statistical inference for short-sample pro-
blems, and examining the impact of outliers. In subsamples, the relation between
excess returns and the political variables remains signi¢cant. However, the level
of signi¢cance drops from 5 to 10 percent, largely because the power of our tests
decreases substantially with the number of elections.We run a bootstrap experi-
ment to correct small-sample inference problems. The corrected statistics corro-
borate the signi¢cance of the relation between political cycles and the equity
premium. Finally, we use quantile regressions to establish that outliers do not
drive our results. Of course, given the limitations of the data, we can never be
absolutely sure that the impact of political cycles on the stock market is not just
a statistical £uke.
We examine whether the di¡erence in average returns is due to a di¡erence in
expected returns or a di¡erence in unexpected returns. In the ¢rst case, the di¡er-
ence in realized returns would be due to a ‘‘Democratic risk premium.’’1 In the
second case, the di¡erence in returns would be driven by surprises in the econom-
ic policies of the party in the presidency. In other words, a di¡erence in unex-
pected returns would occur when the policies enacted by Republicans and
Democrats deviate systematically from what the market anticipates.
To investigate whether the di¡erence in returns was expected or unexpected,
we use three di¡erent approaches. First, the presidential-partisan cycle might
merely be proxying for variations in expected returns due to business cycle £uc-
tuations. Indeed, previous research has found that GDP growth is slower during
Republican presidential mandates, and that Democratic administrations have
been associated with signi¢cantly higher in£ation rates.2 There is also substan-
tial evidence that macroeconomic variables related to the business cycle can fore-
cast stock market returns.3 Therefore, the e¡ect of political variables on the
stock market might only be proxying for variations in the business cycle. To test
this ‘‘proxy’’ hypothesis, we examine the relation between stock market returns
and political variables using macro variables known to forecast the stock market
as controls for business cycle £uctuations. After controlling for the dividend-
price ratio, the default and term spreads, and the relative interest rate, our re-
sults become even stronger. The di¡erence between Democratic and Republican
presidencies is still around 10 percent for value-weighted returns and 20 percent
for equal-weighted returns, statistically signi¢cant, and stable over di¡erent
sample periods. Presidential parties thus capture variations in returns that are
largely uncorrelated to what is explained by business cycle £uctuations.
1
Since we do not observe a di¡erence in volatility between Republican and Democratic
mandates (see Section IV.C below), the risk that could justify a premium for Democrats would
have to be of the ‘‘Peso’’ type, where there would be a (perceived) higher probability of Demo-
crats enacting economic policies detrimental to the stock market.
2
See Alesina and Rosenthal (1995), Alesina, Roubini, and Cohen (1997), and references
therein.
3
See, for example, Chen, Roll, and Ross (1986), Keim and Stambaugh (1986), Campbell and
Shiller (1988), Fama and French (1988, 1989), Campbell (1991), and Fama (1991).Political Cycles and the Stock Market 1843
In a related experiment, we use the same business cycle variables to decompose
returns into expected returns and unexpected returns. We simply regress rea-
lized returns on the business cycle variables and take the ¢tted values of the re-
gression to be expected returns and the regression residuals to be unexpected
returns. We then analyze di¡erences in each component under Republican and
Democratic presidencies.We ¢nd that most of the observed di¡erence in returns
can be attributed to a di¡erence in unexpected returns. For the entire sample, we
¢nd that expected returns are actually 1.8 percent higher under Republicans,
whereas unexpected returns are 10.8 percent higher under Democrats. It thus
seems that the di¡erence in realized returns can be attributed to the market
being systematically positively surprised by Democratic policies. Of course, the
Fama critique applies here.We cannot rule out the possibility that the higher re-
turns under Democrats correspond to compensation for risk since we do not
know for sure what drives the variation in the risk premium. Any test of rational
expectations like the one we implement is simultaneously a test of rational ex-
pectations and of the risk premium.
To further study the di¡erence in returns, we examine whether the relation be-
tween returns and the presidential-partisan cycle is concentrated around election
dates. If the di¡erence in returns was due to a higher ex ante risk premium, we
should observe a large movement in stock prices when the uncertainty about
which party wins the presidency is resolved. We ¢nd no signi¢cant evidence of
stock price changes immediately before, during, or immediately after elections.4
To the contrary, the di¡erence in returns grows gradually over the term of the pre-
sidency.This supports the hypothesis that the di¡erence in returns was not antici-
pated by the market and was due to (systematic) surprises in economic policies.
It can be argued that the timing of the resolution of uncertainty is hard to as-
certain. In fact, the results of most elections are largely anticipated so that it is
di⁄cult to determine when exactly the winner is known. To get around this pro-
blem, we examine the reaction of the stock market to the result of the four most
contested (and hence hardest to predict) presidential elections.We ¢nd no signif-
icant evidence of large returns immediately before or after surprise Republican
or Democratic victories.
As a ¢nal test of the hypothesis that the higher realized returns under Demo-
cratic presidencies might be compensation for risk, we examine whether indeed
risk was any higher under Democrats than Republicans. This di¡erence in riski-
ness might arise from di¡erences in economic policies pursued by each party or
from varying levels of uncertainty among investors about these policies. If there
was indeed a di¡erence in the riskiness of the stock market, it would be reason-
able to argue that it should command a risk premium to compensate investors for
the greater risks incurred in those periods. However, we ¢nd that market volati-
lity is actually higher under Republican presidents, contrary to the hypothesis.5
4
This ¢nding is consistent with the evidence in Cutler, Poterba, and Summers (1989) that
important news is seldom related to large stock market returns, and vice versa.
5
Although, after controlling for the state of the economy, the di¡erence in risk under the
two regimes becomes insigni¢cant.1844 The Journal of Finance
Of course, we are left with the possibility that ‘‘Democratic risk’’ is of the ‘‘Peso’’
type, and we just happened not to observe any bad realization in our sample. Un-
fortunately, that possibility cannot be tested.
Given the results above, we are left with a puzzle. How can such a large and
persistent di¡erence in returns exist in an e⁄cient market if it is not compensa-
tion for risk? We can speculate that the di¡erence in returns is due to di¡erences
in economic policies between Republicans and Democrats. However, to be consis-
tent with the ¢ndings, these policies must impact the stock market directly and
not just through their e¡ect on the state of the economy. It is not immediately clear
what kind of policies can have this e¡ect. Second, di¡erences in economic policy
may justify our results only if they were unexpected by the market. In other words,
under this explanation, market participants must have been systematically posi-
tively surprised by Democratic policies. The obvious question is then, why have
investors not learned about the di¡erence in party policies and adjusted stock
prices when the result of the election becomes known? We cannot provide a con-
clusive answer to this question.We can only conjecture that investors perceive the
party in the presidency to be a noisy signal of economic policy. Moreover, given the
small number of presidencies, it may have been di⁄cult for investors to learn
about systematic di¡erences in policies. Until we ¢nd answers to these questions,
the relation between the political cycle and the stock market remains a puzzle.
A clear possibility is that our ¢ndings might be the product of data mining.
Taking into account that, over the years, researchers (and investors) have tried
countless variables to forecast stock market returns, it might just be the case that
we have stumbled upon a variable that tests signi¢cantly even when there is actu-
ally no underlying relation between the presidency and the stock market. As
pointed out by many authors, and illustrated by Sullivan, Timmermann, and
White (2001), if one correlates enough variables with market returns, some spur-
ious relations are likely to be found. The possibility of data mining is certainly a
concern in the case of the presidential party variable. Indeed, we have tried other
political variables, related to the party in control of Congress, without success.
Additionally, our empirical investigation is not preceeded by a clear theoretical
model; it is only motivated by a conjecture. One way to address data mining is to
use the Bonferroni approach and adjust the con¢dence level of the tests by the
number of hypotheses tested. In our case, taking into account that we also looked
at another political institution besides the presidency, we should double the p-va-
lues of the tests. If we do that, some of the hypotheses we test are no longer sig-
ni¢cant. Unfortunately, using the Bonferroni approach severely reduces the
power of the tests.The reduction in power is of particular hindrance in our study
since, as we show below, our tests already have modest power. Finally, in defense
of the robustness of our ¢ndings, we should point out that politics, unlike ‘‘butter
production in Bangladesh,’’6 is known to have a pervasive impact on the economy.7
6
Leinweber (1997) searched through a United Nations database and discovered that, his-
torically, the single best predictor of the Standard & Poor’s 500 stock index was butter pro-
duction in Bangladesh.
7
See Alesina and Rosenthal (1995), Alesina et al. (1997), and references therein.Political Cycles and the Stock Market 1845
Other authors have documented the di¡erence in stock returns under Repub-
lican and Democratic presidents, notably Herbst and Slinkman (1984), Huang
(1985), Hensel and Ziemba (1995), Siegel (1998), and Chittenden, Jensen, and John-
son (1999). Our paper is the ¢rst to formally test the relation between political
cycles and the stock market, examine the robustness of this relationship, investi-
gate cross-sectional returns, and use macroeconomic control variables. There is
also a rich empirical and theoretical literature about the e¡ects of political cy-
cles on the macroeconomy. For surveys in this area, see Alesina et al. (1997) and
Drazen (2000).These books o¡er convincing evidence that political variables have
an impact on the state of the macroeconomy. Some of our tests are loosely moti-
vated by hypotheses formulated in that literature.
The rest of the paper is structured as follows. Section I introduces the
data and the notation used in the paper. Section II discusses the empirical
methods and presents the main results: the signi¢cant and robust correlation
between excess market returns and presidential-partisan variables. Section III
investigates whether the results are spurious with a battery of robustness
tests and discusses the possibility of data mining. In Section IV, we test
three hypotheses concerning the di¡erences in returns across political cycles
and establish that the di¡erence in returns across political regimes was not
expected by investors. Section V sets out the research agenda for future work
and concludes.
I. Data
In this section, we describe the variables used in the study. For clarity of
exposition, the data are categorized into ¢nancial variables, political variables,
and control variables. Table I provides summary statistics for quick reference.
All series are at monthly frequency. The entire sample period, 1927:01^1998:12,
contains 864 monthly observations, 18 elections, 10 Democratic and 8 Republican
presidents.8 As a check of robustness, we perform the statistical analysis on the
full sample and two equal subsamples. The ¢rst subsample, 1927:01^1962:12, in-
cludes the Great Depression, the subsequent recovery, and World War II. It con-
tains 432 observations and spans three Republican and six Democratic
presidencies. The second subsample, from 1963:01 to 1998:12, covers the most re-
cent period and includes 432 months under ¢ve Republican and six Democratic
presidents.
A. Financial Variables
We use the log monthly returns of the value-weighted (VWRt) and equal-
weighted (EWRt) portfolios from CRSP. The log interest rate (TBLt) is computed
from the three-month Treasury bill, obtained from Ibbotson Associates. INFt is
the log monthly in£ation, also from Ibbotson Associates. Additionally, we use
8
The sample starts in 1927 whereas CRSP o¡ers return data since 1926. We lose one year of
data to be able to run regressions with control variables that involve lagged data.1846 The Journal of Finance
Table I
Summary Statistics of Financial and Control Variables
The table reports the sample average (Mean), standard deviation (Std.Dev.), and the autoregres-
sive coe⁄cient (A.R.) of all ¢nancial series and control variables used in this study. All returns
are computed in logarithmic form and expressed in annualized percentage points.
1927:01^1998:12 (864 obs) 1927:01^1962:12 (432 obs) 1963:01^1998:12 (432 obs)
Series Mean Std.Dev. A.R. Mean Std.Dev. A.R. Mean Std.Dev. A.R.
VWR-TBL 6.46 19.20 0.20 7.49 22.44 0.30 5.42 15.32 0.01
VWR-INF 7.08 19.20 0.17 7.29 22.38 0.24 6.88 15.41 0.07
EWR-TBL 8.76 25.32 0.25 10.51 29.99 0.29 7.02 19.58 0.22
EWR-INF 9.39 25.25 0.22 10.31 29.89 0.25 8.47 19.58 0.22
TBL-INF 0.60 1.94 0.82 0.26 2.53 0.80 1.46 1.01 0.88
VOL 15.59 0.56 0.87 17.63 0.68 0.90 13.56 0.38 0.72
DEC1-TBL 8.43 3.29 0.27 11.25 4.13 0.27 5.61 2.16 0.30
DEC2 -TBL 7.24 2.96 0.29 8.99 3.65 0.33 5.49 2.07 0.21
DEC3 -TBL 7.77 2.75 0.22 9.18 3.32 0.29 6.37 2.02 0.07
DEC4 -TBL 7.77 2.56 0.22 9.07 3.06 0.29 6.48 1.95 0.07
DEC5 -TBL 7.52 2.51 0.16 8.58 3.02 0.23 6.45 1.87 0.03
DEC6 -TBL 7.75 2.41 0.22 9.55 2.89 0.30 5.95 1.81 0.02
DEC7-TBL 6.95 2.31 0.17 8.38 2.74 0.25 5.52 1.78 0.03
DEC8-TBL 7.11 2.16 0.16 8.17 2.54 0.28 6.04 1.70 0.15
DEC9 -TBL 6.95 2.06 0.20 8.59 2.44 0.31 5.31 1.60 0.10
DEC10 -TBL 6.38 1.81 0.22 7.16 2.12 0.29 5.60 1.43 0.06
DP 3.07 0.33 0.98 2.90 0.27 0.95 3.24 0.29 0.99
DSP 1.14 0.02 0.97 1.27 0.03 0.97 1.01 0.01 0.97
TSP 1.64 0.04 0.91 1.61 0.03 0.94 1.67 0.04 0.90
INF 3.08 0.19 0.83 1.50 0.24 0.76 4.67 0.11 0.91
RR 0.01 0.03 0.74 0.01 0.02 0.61 0.03 0.04 0.77
cross-sectional returns from 10 size decile portfolios (DECjt, for j ¼ 1, 2,y,10),
obtained from Kenneth French.We conduct the statistical analysis in this paper
with excess and real returns; for example, when studying the value-weighted
portfolio, we compute VWRt TBLt (log value-weighted return minus log inter-
est rate) and VWRt INFt (log value-weighted return minus log in£ation).9 We
compute the monthly volatility of the value-weighted portfolio return (VOLt)
from within-month daily return data, using the approach of French, Schwert,
and Stambaugh (1987). The daily return data is from Schwert (1990).
Although there are (limited) return and interest rate series going further back
in time (Schwert (1990)), two main reasons lead us to restrict the analysis to the
post-1927 period. First, there is evidence that the ideologies of the Democratic
9
This is the most convenient way to abstract from the e¡ects of in£ation and monetary
policy. Political macroeconomists have widely agreed that in£ation is higher during Demo-
cratic terms. Fama (1981), Geske and Roll (1983), Kaul (1987), and Goto and Valkanov (2000)
provide evidence of the e¡ect of monetary policy on returns and in£ation.Political Cycles and the Stock Market 1847
and Republican parties beforeWWI were not clearly delineated. Second, the data
for most of the control variables are not available prior to 1927.
B. Political Variables
We de¢ne the following presidential cycle dummy variables:
RDt ¼ 1 if a Republican is in o⁄ce at time t; RDt ¼ 0 otherwise.
DDt ¼ 1 if a Democrat is in o⁄ce at time t; DDt ¼ 0 otherwise.
The political index variable that we use is motivated by previous political
macroeconomic studies. It can be motivated by a ‘‘partisan’’ view of political cy-
cles discussed in Hibbs (1977) and Alesina (1987), which emphasizes the di¡ering
motivations and political platforms of the political parties.10 For instance, this
school argues that policies related to corporate, personal income, and consump-
tion taxes, government spending, insurance coverage, and social bene¢ts are dif-
ferent under Republicans and Democrats.
We also studied the impact of Congressional variables on the stock market.
Since we found no signi¢cant relation, we do not present these results in the pa-
per. They are, however, available upon request.
C. Control Variables
The conditioning variables we use are the annualized log dividend-price ratio
(DPt), the term spread (TSPt) between the yield to maturity of a 10 -year Treasury
note and the three-month Treasury bill, the default spread (DSPt) between yields
of BAA- and AAA-rated bonds, and the relative interest rate (RRt) computed as
the deviation of the three-month Treasury bill rate from its one-year moving aver-
age. The dividend price ratio is from CRSP, whereas the other conditioning vari-
ables are from the DRI database.
We have tried to be as exhaustive in our list of conditioning variables as possi-
ble. Some of these variables may be correlated. However, if we had to err, we
wanted to err on the side of including redundant information, rather than forget-
ting relevant information which would lead to inconsistent estimates. The use of
these control variables is uncontroversial, as they all have been used more than
once in previous studies. Some of the most widely cited papers that take the divi-
dend-price ratio, the term spread, or the default spread as predictors are Keim
and Stambaugh (1986), Campbell and Shiller (1988), Fama and French (1988,
1989), and Fama (1991).The power of the relative interest rate to forecast expected
returns was argued by Campbell (1991) and Hodrick (1992).
10
We have also experimented with index variables that allow us to test for abnormal re-
turns before or at any time during the election period, irrespective of the political party in
power. Such variables can be motivated by ‘‘opportunistic’’ models of political behavior, where
policymakers choose policies that maximize their chances of staying in o⁄ce regardless of
their own party’s political platform. See Nordhaus (1975), Lindbeck (1976), Rogo¡ (1990), and
Persson and Tabellini (1990). It will become clear in the subsequent analysis that such ‘‘oppor-
tunistic’’ models are not supported by the data, as we do not observe large price movements
around elections.1848 The Journal of Finance
II. Main Finding
In this section, we establish the empirical fact that presidential-partisan cycles
have been associated with returns in the stock market as well as with the real
risk-free interest rate. Furthermore, we document that the di¡erence in stock
market returns under Republican and Democratic presidencies is robust in dif-
ferent subsamples.
Figure 1 plots the average excess value-weighted annual return during each
presidency in the 1927 to 1998 period. Republican periods are shaded in a darker
color and the dash-dotted line denotes the unconditional mean of the series. Ex-
cess returns under Republican presidents have been historically lower than un-
der Democratic presidents. Only 1 (out of 10) Democratic presidency (Roosevelt,
1937 to 1941) has known signi¢cantly lower than average excess returns, and only
30
20
Roosevelt (D): 1937−1941
Hoover (R): 1929 −1933
Excess Value Weighted Returns
10
0
Roosevelt/Truman (D): 1941−1945
Kennedy/Johnson (D): 1961−1965
Eisenhower (R): 1953−1957
Eisenhower (R): 1957−1961
Nixon/Ford (R): 1973−1977
Roosevelt (D): 1933 −1937
Johnson (D): 1965−1969
Truman (D): 1949 −1953
Truman (D): 1945−1949
Reagan (R): 1981−1985
Reagan (R): 1985−1989
-10
Clinton (D): 1993−1997
Clinton (D): 1997−1998
Carter (D): 1977−1981
Nixon (R): 1969 −1973
Bush (R): 1989−1993
-20
-30
1930 1940 1950 1960 1970 1980 1990
Years
Figure 1. Average annual excess returns by presidential term, 1927 to 1998. Figure 1
displays the average annualized excess value-weighted returns during each presidential
term for the 1927 to 1998 period. Republican administrations are denoted with a darker
shade. The average excess return through the entire sample is marked as a dash-dotted
line. Most Democratic presidencies have been associated with higher than average excess
returns, with Roosevelt’s (1937^1941) tenure being the only signi¢cant exception. Simi-
larly, most Republican presidencies have been associated with signi¢cantly lower than
average returns, with the only exception Eisenhower (1953 ^1957).Political Cycles and the Stock Market 1849
1 (out of 8) Republican presidency (Eisenhower, 1953 to 1957) has been associated
with signi¢cantly higher than average returns.
To measure the correlation between (excess and real) returns and political
variables, we run the following regressions:
rtþ1 ¼ a þ bpt þ utþ1 ð1Þ
where returns are denoted by rt þ 1 and the political variable by pt. The timing of the
variables emphasizes that the political variables are known at the start of the return
period. Under the null hypothesis of political cycles having no e¡ect on returns, we
should have b ¼ 0 in the regression.Table II, Panel A, presents the results from regres-
sing the excess and real returns of the value-weighted and equal-weighted portfolios
and for the Treasury bill on index variables for Republican and Democratic presiden-
tial mandates.11 The coe⁄cients are simply the means of returns during the Republi-
can and Democratic presidencies. The probabilities of accepting the null hypothesis
(p-values) reported below the estimates are computed using asymptotic standard er-
rors using the Newey^West (1987) approach to correct for heteroskedasticity and seri-
al-correlation, as well as using bootstrapped standard errors obtained by resampling
the residuals of the regressions. If the residuals are conditionally heteroskedastic, the
¢nite-sample distribution of the t-statistics is better approximated by the bootstrap.
However, the bootstrap is not appropriate if the residuals are serially correlated, nor
does it correct for spurious correlation between the returns and the political vari-
ables. For that, we need to bootstrap the regressor pt itself, as we do in the next
section. All results are presented for the entire sample and for the two subsamples.
During the 1927 to 1998 period, the value-weighted excess return under a
Democratic White House was 10.69 percent per year, whereas it was only 1.69 per-
cent per year under a Republican president, amounting to a di¡erence of 9.01 per-
centage points, which is economically and statistically signi¢cant. It is
interesting to notice that the di¡erence in excess returns is due to both the real
stock market return being higher and the real Treasury bill rate being lower un-
der Democrats than Republicans.12 For the full sample, the 9.01 percent di¡erence
in excess return of the value-weighted index can be decomposed into a higher
average stock market return of 5.31 percent under Democrats and the real T-bill
rate being 3.70 percent lower under Democrats.
It is remarkable that the di¡erence in returns is found robustly in the two sub-
samples.13 In the 1963 to 1998 period, which is the most favorable for Republicans,
11
We actually run a regression of market returns on Republican (RD) and Democratic (DD)
presidential dummies, or rt þ 1 ¼ a1RDt þ a2DDt þ ut þ 1. The hypothesis of no di¡erence between
the coe⁄cients, or a1 a2 ¼ 0, is equivalent to b ¼ 0 in regression (1).
12
This shows the importance of using excess returns to test the correlation with political
variables. In contrast, previous studies have concentrated on stock returns rather than excess
returns, and ¢nd smaller di¡erences between Republican and Democratic administrations,
generally on the order of ¢ve percent (Hensel and Ziemba (1995), Chittenden et al. (1999),
and Siegel (1998)). All those studies used the S&P 500 Index as a proxy for the stock market.
13
We tried a variety of other subsample schemes, being only constrained by the need to
ensure that each subsample contains a su⁄ciently large number of months when each party
was in the presidency. The results were always similar to the numbers we report.1850 The Journal of Finance
the di¡erence is 6.85 percent. The magnitude of this di¡erence is still highly sig-
ni¢cant in economic terms.The p-values of the di¡erence in mean returns for this
subsample are 0.07 and 0.09. For the ¢rst subsample, going from 1927 to 1962, the
di¡erence in returns is 9.45 percent, which is highly signi¢cant in both economic
and statistical terms.
Obtaining statistical signi¢cance in subsamples is surprising given the
low power of our test, especially in periods during which only a few presidential
elections were held. In a Monte Carlo exercise, we simulated the power of our
test, for the given number of Republican and Democratic presidents and
number of observations in each sample.The simulation experiment assumes that:
(1) there is a di¡erence of nine percent between the two parties, (2) the distribu-
tion of the residuals is the same as the sample data, and (3) the proportion
of Republican presidencies is the same as in each subsample. Each simulated
sample (of the same length as each subsample) is generated from the following
process:
rtþ1 ¼ a1 RDt þ a2 DDt þ utþ1 ð2Þ
where a1 ¼ 2, a2 ¼ 11. The presidential dummy variables are simulated according
to their frequency in the subsample and ut þ 1 is bootstrapped from the data. Un-
der the null of our tests, there is no di¡erence in returns, so a1 ¼ a2. Here, we are
interested in the power of the test for a1 ¼ a2 versus the particular alternative,
a2 a1 ¼ 9. For the entire sample, the power to reject the null hypothesis (of zero
di¡erence) at the ¢ve percent signi¢cance level, when the di¡erence between the
means in returns is in fact nine percent, is only 0.51. The low power of the test is
mostly due to the high volatility of returns.The power decreases to 0.23 in the ¢rst
subsample and to 0.28 in the second subsample. The lower power in the subsam-
ples is due to the smaller number of observations and the fact that the two parties
are not evenly distributed in the subsamples. In particular, the ¢rst subsample
only includes three Republican administrations, corresponding to 179 out of 431
months, or roughly 40 percent of the sample. Moreover, the lower power of the
test in the ¢rst subperiod (1927 to 1962) is due to the higher variance in returns
compared to the variance of returns in the second subsample during that period
(see Table I).
The di¡erence in excess returns of the equal-weighted portfolio in Table II,
Panel A, is even more dramatic, 16.52 percent in the full sample and 14.93 and
17.19 percent in the subsamples. This di¡erence is mostly due to the much higher
average real return on this portfolio under Democratic administrations. This re-
sult implies a di¡erential e¡ect of the political cycle on small and large ¢rms.We
investigate this size e¡ect more closely by examining the returns to 10 size-
decile portfolios.When we regress the excess returns of the 10 size-decile portfo-
lios on the presidential index variables, we observe that the di¡erence between
returns is perfectly inversely related to the market capitalization of the compa-
nies. The smallest companies (decile 1) display the largest disparity in returns
during the Republican and Democratic presidencies, 21 percent per year, during
the entire sample period. The di¡erence in returns of the biggest com-
panies (decile 10) remains economically and statistically signi¢cant, 7.71 percent,Political Cycles and the Stock Market 1851
Table II
Average Returns under Republican and Democratic Presidents
Panel A reports mean excess and real returns of value-weighted and equal-weighted portfolios,
VWR-TBL,VWR-INF, EWR-TBL, EWR-INF, and the real interest rate,TBL-INF, during Repub-
lican (RD) and Democratic (DD) presidential terms. All rates are represented in annualized
percentage points. The numbers below the coe⁄cients in the RD and DD columns represent p-
values under the null hypothesis that the estimates are not signi¢cantly di¡erent from zero.The
¢rst number is the p-value of the test conducted using Newey^West (1987) heteroskedasticity
and serial-correlation robust t-statistics.The second number is the p-value of the test conducted
using a conditional bootstrap t-statistic.The p-values below the coe⁄cients in the ‘‘Di¡’’column
are obtained from the Newey^West and conditional bootstrap t-statistics under the null that
there is no di¡erence in returns during Republican and Democratic regimes.The row ‘‘T/Repub-
licans’’displays the number of observations and the number of months of Republican adminis-
trations during the estimation period.The row ‘‘R 2’’displays the average adjusted R2 obtained in
the regressions. Panel B reports the results from a robustness exercise, designed to test whether
the results obtained in Panel A might be due to small sample biases. The maintained null hy-
pothesis is of no relation between returns and political variables. To ¢nd the small sample dis-
tribution of the t-statistic under the null, we draw 10,000 samples of T observations of the
political variables series independently of the return series. For each sample, we obtain the
bootstrapped t-statistics of interest and use their distribution across samples to compute
p-values of the estimates. The numbers in square brackets are the estimates obtained from this
randomization-bootstrap.The next line contains the p-values from the randomization-bootstrap.
1927:01^1998:12 1927:01 1962:12 1963:01 1998:12
RD DD Di¡ RD DD Di¡ RD DD Di¡
Panel A: Signi¢cance Tests
VWR-TBL 1.69 10.69 9.01 1.68 11.13 0.45 2.60 9.45 6.85
0.33 0.00 0.03 0.40 0.01 0.06 0.23 0.00 0.07
0.31 0.01 0.02 0.39 0.02 0.04 0.20 0.03 0.09
VWR-INF 4.25 9.56 5.31 5.22 8.54 3.32 4.50 10.21 5.71
0.12 0.00 0.13 0.22 0.03 0.17 0.10 0.00 0.12
0.13 0.00 0.13 0.17 0.06 0.16 0.10 0.01 0.13
EWR-TBL 0.01 16.52 16.52 1.30 16.23 14.93 0.02 17.21 17.19
0.50 0.00 0.01 0.44 0.00 0.04 0.50 0.00 0.01
0.46 0.00 0.01 0.45 0.00 0.03 0.48 0.00 0.01
EWR-INF 2.58 15.38 12.80 4.84 13.63 8.79 1.94 17.95 16.00
0.29 0.00 0.02 0.28 0.02 0.10 0.33 0.00 0.01
0.29 0.00 0.03 0.32 0.02 0.08 0.31 0.00 0.01
TBL-INF 2.54 1.16 3.70 3.50 2.66 6.16 1.89 0.79 1.10
0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.02
0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.01 0.00
T/Republicans 863/407 431/179 431/239
2
R 0.01 0.01 0.011852 The Journal of Finance
Table II
(Continued)
1927:01^1998:12 1927:01 1962:12 1963:01 1998:12
RD DD Di¡ RD DD Di¡ RD DD Di¡
Panel B: Robustness Tests
VWR-TBL 1.69 10.69 9.01 1.68 11.13 9.45 2.60 9.45 6.85
[6.46] [6.46] [0.00] [7.49] [7.49] [0.00] [5.42] [5.42] [0.00]
0.04 0.03 0.04 0.10 0.16 0.09 0.09 0.05 0.05
VWR-INF 4.25 9.56 5.31 5.22 8.54 3.32 4.50 10.21 5.71
[7.08] [7.08] [0.00] [7.29] [7.29] [0.00] [6.88] [6.88] [0.00]
0.11 0.17 0.20 0.28 0.36 0.20 0.22 0.17 0.15
EWR-TBL 0.01 16.52 16.52 1.30 16.23 14.93 0.02 17.21 17.19
[8.76] [8.76] [0.00] [10.51] [10.51] [0.00] [7.02] [7.02] [0.00]
0.02 0.01 0.01 0.07 0.13 0.08 0.04 0.01 0.01
EWR-INF 2.58 15.38 12.80 4.84 13.63 8.79 1.94 17.95 16.00
[9.39] [9.39] [0.00] [10.31] [10.31] [0.00] [8.47] [8.47] [0.00]
0.03 0.03 0.04 0.13 0.15 0.09 0.04 0.02 0.01
TBL-INF 2.54 1.16 3.70 3.50 2.66 6.16 1.89 0.79 1.10
[0.60] [0.60] [0.00] [ 0.26] [ 0.26] [0.00] [1.46] [1.46] [0.00]
0.01 0.04 0.01 0.01 0.05 0.01 0.24 0.22 0.05
but is three times smaller. The results from the subsamples are very similar.
This ¢nding explains the di¡erence between the results in the value-
weighted and equally weighted regressions. The former put (relatively) more
weight on large companies, whereas the latter put more weight on small
companies.
It could be argued that the di¡erences in the e¡ect of political variables on the
excess returns of the size-decile portfolios is simply due to the fact that small
stocks tend to have higher betas on the market than big stocks. In that case, po-
litical variables would only a¡ect the overall level of the market and the large
e¡ect on small stocks is due to their high sensitivity to market moves. To investi-
gate this possibility, we run regressions of the excess returns of the size-decile
portfolios on the political variables together with the excess return on the va-
lue-weighted portfolio. Table III shows the estimates of the coe⁄cients on RD
and DD.The betas of the size deciles portfolios (not displayed in the table for brev-
ity) vary from 1.39 for the smallest companies to 0.93 for the largest ones. We see
that, after controlling for the di¡erences in market beta, the political variables
retain considerable explanatory power for the di¡erence in expected returns of
portfolios formed according to size. The di¡erence for the smallest decile is still
on the order of 10 to 15 percent. For the overall sample, the di¡erence in ‘‘beta-
adjusted’’ mean returns is signi¢cant for all size-decile portfolios. In the more
recent subsamples, the statistical signi¢cance disappears for the biggest ¢rms
but remains high for smaller stocks. It thus seems that the ‘‘size e¡ect’’ is asso-
ciated with the political cycle.Political Cycles and the Stock Market 1853
Table III
Average Returns of Size-Decile Returns under Republican and
Democratic Presidents, Controlling for Market Returns and Di¡erences
in Market Betas
Table III reports the results from the regression: DEC( j)t+1 TBLt+1 ¼ a1, jRDt þ a2,jDDt þ
bj(VWRt+1 TBLt) þ et+1, estimated for three sample periods and j ¼ 1,y,10. The estimates of b
are omitted, for clarity of exposition. They range from 1.39 for the Decile 1 to 0.93 for Decile 10
and are relatively stable across samples. The numbers below the coe⁄cients in the RD and DD
columms represent p-values of a t-test under the null hypothesis that the estimates are not sig-
ni¢cantly di¡erent from zero.The ¢rst number is the p-value of the test conducted using Newey^
West (1987) heteroskedasticity and serial-correlation robust t-statistics. The second number is
the p-value of the test conducted using a conditional bootstrap t-statistic. The p-values below
the coe⁄cients in the ‘‘Di¡’’ column are obtained from the Newey^West and conditional boot-
strap t-statistics under the null that there is no di¡erence in returns during Republican and
Democratic regimes, or a1 ¼ a2.
Signi¢cance Tests
1927:01^1998:12 1927:01^1962:12 1963:01^1998:12
RD DD Di¡ RD DD Di¡ RD DD Di¡
DEC1-TBL 1.91 14.18 12.27 7.81 13.39 5.58 1.25 15.47 16.72
(small) 0.25 0.00 0.00 0.08 0.00 0.23 0.31 0.00 0.00
0.27 0.00 0.00 0.12 0.00 0.24 0.36 0.00 0.00
DEC2 -TBL 2.13 11.71 9.57 4.86 11.47 6.61 0.82 11.99 11.16
0.16 0.00 0.00 0.13 0.00 0.12 0.34 0.00 0.00
0.20 0.00 0.00 0.18 0.00 0.13 0.33 0.00 0.00
DEC3 -TBL 3.26 11.75 8.49 5.01 11.73 6.72 2.74 11.37 8.63
0.03 0.00 0.00 0.07 0.00 0.06 0.04 0.00 0.00
0.03 0.00 0.00 0.09 0.00 0.05 0.08 0.00 0.00
DEC4 -TBL 4.12 10.98 6.85 5.73 11.11 5.38 3.36 10.76 7.40
0.00 0.00 0.00 0.03 0.00 0.08 0.01 0.00 0.01
0.01 0.00 0.00 0.02 0.00 0.08 0.02 0.00 0.01
DEC5 -TBL 4.09 10.57 6.48 5.21 10.71 5.50 3.86 9.93 6.06
0.00 0.00 0.00 0.03 0.00 0.05 0.00 0.00 0.01
0.00 0.00 0.00 0.02 0.00 0.03 0.00 0.00 0.00
DEC6 -TBL 5.37 9.84 4.47 7.31 10.91 3.60 4.31 8.10 3.78
0.00 0.00 0.01 0.00 0.00 0.09 0.00 0.00 0.07
0.00 0.00 0.01 0.00 0.00 0.08 0.00 0.00 0.03
DEC7-TBL 4.97 8.68 3.71 6.30 9.63 3.33 4.54 6.64 2.09
0.00 0.00 0.01 0.00 0.00 0.08 0.00 0.00 0.16
0.00 0.00 0.00 0.00 0.00 0.07 0.00 0.00 0.12
DEC8 -TBL 5.51 8.51 3.00 6.63 9.13 2.50 5.27 6.88 1.61
0.00 0.00 0.01 0.00 0.00 0.13 0.00 0.00 0.16
0.00 0.00 0.00 0.00 0.00 0.08 0.00 0.00 0.11
DEC9 -TBL 5.49 8.25 2.76 7.02 9.58 2.57 4.99 5.49 0.49
0.00 0.00 0.00 0.00 0.00 0.05 0.00 0.00 0.34
0.00 0.00 0.00 0.00 0.00 0.03 0.00 0.00 0.33
DEC10 -TBL 5.49 7.20 1.71 6.62 7.55 0.93 5.32 5.76 0.43
(big) 0.00 0.00 0.02 0.00 0.00 0.12 0.00 0.00 0.37
0.00 0.00 0.01 0.00 0.00 0.12 0.00 0.00 0.401854 The Journal of Finance
III. Are the Results Spurious?
The di¡erence in returns is intriguing not only because of its economic signif-
icance, but also because it is so stable across subperiods. However, there is a
possibility that the ¢ndings are driven by a few outliers or that our statis-
ical inference is plagued by small-sample biases. We have to acknowledge the
possibility that the correlation between returns and the political variables is
spurious, especially since our tests are not motivated by a theoretical model. In
the following subsections, we use a randomization-bootstrap procedure and
quantile regressions to demonstrate that the observed di¡erence in returns dur-
ing Republican and Democratic administrations is a robust feature of the data.
A. A Randomization-Bootstrap Experiment
The standard errors in Table II, Panel A are robust to serial correlation and het-
eroskedasticity in the residuals. However, the results may still be driven by a‘‘lucky
draw’’ from the political variables. After all, there are only 18 presidencies in our
sample and even fewer switches of theWhite House between parties.To address this
concern, we turn to a randomization-bootstrap procedure,14 which is formally devel-
oped in Davison and Hinkley (1977) and Efron and Tibshirani (1998).15 We address
the following question: How likely would it be to observe such a di¡erence in returns
across political regimes if the regimes were truly independent of returns?
To ¢nd the small-sample distribution of the t-statistic, t, under the null, we con-
duct the following resampling experiment.We draw samples of Tobservations each
by keeping the series of returns as is and drawing the political variables indepen-
dently from the returns. This resampling is done in such a way as to be consistent
with the dates of presidential changes.We produce 10,000 time series of frtþ1 ; pt gTt¼1
for which there is no relation between returns and the party in the presidency. De-
note the jth sample by frjtþ1 ; pjt gTt¼1, for j ¼ 1,y,10,000. We can compute b ^ j and the
j
corresponding t as in regression (1). The bootstrapped distribution of t under
the null hypothesis is simply the distribution of the 10,000 draws of t j. The mean of
^ is denoted by b ^
the bootstrapped distribution of b . Under the null, returns during
Democratic and Republican presidencies must be equal to each other and to the
^ ¼ 0. The two-sided bootstrapped p-va-
unconditional mean, which implies that b
lue is computed as pboot ¼ (#{t t} þ #{t j t}/10,000) where #{tj t} denotes
j
the number of bootstrapped tj’s that are higher than the computed t statistic.16
Table II, Panel B, presents the results from the randomization-bootstrap tests.
The ¢rst number (in square brackets) below the estimates is the mean of the
14
We thank the referee for this excellent suggestion.
15
There are several equivalent ways of setting up this bootstrap. We chose a setup that lends
itself to a multivariate generalization, which allows us to extend the analysis with control
variables in the regressions.
16
The bootstrap experiment can be carried out by bootstrapping either the distribution of b ^
or the distribution of t ¼ b ^=seðb
^Þ . In our application, both methods yield almost identical
results. Bootstrapping the t-statistic is advocated in the bootstrap literature (see Efron and
Tibshirani (1998, pp. 161 and 321) and Davison and Hinkley (1997, p. 268)) since its distribution
is pivotal, that is, it does not depend on (nuisance) parameters.Political Cycles and the Stock Market 1855 corresponding parameter from the randomized samples. As noted above, under the assumption that returns are independent of the political variables, the mean returns under the two regimes should be equal to each other and to the uncondi- tional mean (in Table I). We present this number as a check of the bootstrap pro- cedure and to show that there is little simulation noise in the distribution. The second number below the estimates is the p-value, pboot, under the null that the estimated value of the parameter is equal to the bootstrapped value in square brackets.We focus our attention on the column ‘‘Di¡,’’ which can be compared di- rectly with column ‘‘Di¡’’ in Panel A, since both columns test the hypothesis of no di¡erence in returns between the two regimes. We ¢nd that the di¡erence be- tween political regimes is still mostly signi¢cant at the ¢ve percent level for both the value-weighted and the equal-weighted returns across periods. For the over- all sample, the p-value of the di¡erence is 0.04, which is only marginally higher than the asymptotic p-value of 0.03. The results strongly support the ¢nding that market returns are higher under Democratic than Republican presidencies. In the subsamples, the randomization-bootstrapped p-values di¡er more from the p-values presented in Panel A. For the ¢rst subsample, the p-value is now higher (0.09 vs. 0.04), whereas for the second subsample it is actually lower (0.05 vs. 0.09), which indicates that the previous corrections for serial correlation and hetero- skedasticity may be sometimes too conservative. With three di¡erent p-values, we are faced with the question of which numbers to believe. Rather than choosing arbitrarily one statistical method over another, it is more prudent (and conservative) to consider all results and to take the max- imum p-value in each test. In the entire sample, the di¡erence is statistically sig- ni¢cant at the 5 percent level using any of the testing procedures. In the subsamples, the di¡erence is signi¢cant at the 10 percent level in all tests. We want to emphasize that, while the above resampling procedures are as ex- haustive as we could design them, they are still mere statistical procedures. There is always the possibility that the di¡erence between political parties is spurious. After all, we only have a small number of presidencies, so, no matter how astute the tests are, the limitations of the data are severely binding. B. Quantile Regressions A related concern is whether the results are driven by a few outliers, such as the extremely negative returns during the Hoover administration and/or the unu- sually high returns during the Roosevelt years. To address this concern, we run quantile regressions. We ask whether a particular quantile of the distribution of returns accounts for the di¡erence between Republican and Democratic admin- istrations. Conditional quantiles can be thought of as the inverse of the condi- tional distribution, and therefore contain the same information. By analyzing the entire distribution of returns under the two regimes, we can precisely ¢nd what quantiles of the distribution account for the di¡erence in means. Before discussing our results, we present a brief introduction to quantile re- gressions. Let the unconditional distribution of rt þ 1 be Frtþ1 ðrÞ ¼ Prðrtþ1 rÞ. Then, for any quantile t, 0oto1, we can de¢ne the inverse of Frtþ1 ðÞ as
1856 The Journal of Finance
Qrtþ1 ðtÞ ¼ inffr : Frtþ1 ðrÞ tg. The function Qrtþ1 ð:Þ is called the unconditional
quantile function of rt þ 1. Qrtþ1 ð0:5Þ is the 50th quantile, or the median, of rt þ 1.
The introduction of conditional quantiles is easily understood by making an ana-
logy to the familiar least squares estimation. The conditional mean function
Eðrtþ1 jz ¼ zt Þ ¼ z0t Z, for some explanatory variables z, is estimated by solving
Z ¼ arg minZ St ðrtþ1 z0t ZÞ2 . Similarly, the conditional quantile function
^
Qrtþ1 jz ðtjZ ¼ zt Þ ¼ z0t ZðtÞ can be estimated by solving
X
ZðtÞ ¼ arg min
^ rt ðrtþ1 z0t ZÞ ð3Þ
Z
t
where rt( ) is a piecewise linear ‘‘check function,’’de¢ned as rt(u) ¼ u(t I(uo0))
and I( ) is the indicator function. The function rt(.) selects the quantile t to be
estimated (see Koenker and Hallock (2000)). As above, for the case t ¼ 0.5,
r0.5(u) ¼ juj and the solution of the above problem, ^Zð0:5Þ, is equivalent to minimiz-
ing the sum of absolute values of the residuals. From the de¢nition, Q ^ r jz ð0:5jz ¼
tþ1 t
0^
zt Þ ¼ zt Zð0:5Þ represents the estimate of the conditional median of rt þ 1. For dif-
ferent values of t, the estimate ^ ZðtÞ is the e¡ect of zt on the tth quantile of rt þ 1. An
estimate of the entire function Q ^ r jz ðtjz ¼ zt Þ can be computed from the above
tþ1 t
relation. For a more detailed introduction to quantile regressions, please refer to
Koenker and Hallock (2000) and Koenker (2000).
We run the quantile regression
X
½^ ^ðtÞ ¼ arg min
aðtÞ; b rt ðrtþ1 a bRDt Þ ð4Þ
a;b
t
where the coe⁄cient b(t) captures the quantiles of the di¡erence in returns
between Republicans and Democrats. The estimation is conducted for t ¼ 0.02,
0.04,y,0.98. The results of this quantile regression are plotted in Figure 2, where
the di¡erence b ^ðtÞ is plotted as a solid line, the 95 percent bootstrapped con¢-
dence intervals are plotted in light dashed lines, and the overall unconditional
mean of the di¡erence is plotted for reference. We can clearly see that the di¡er-
ence in returns between Republicans and Democrats is signi¢cant (outside the
con¢dence interval) for quantiles 30 to 60 for value-weighted excess returns and
for quantiles 20 to 75 for equal-weighted excess returns. To summarize, extreme
realizations at the tails of the distribution do not account for the observed di¡er-
ence in returns between Republicans and Democrats.
After the extensive battery of tests applied, we are convinced that the relation
between returns and the political cycle is robust.
C. Data Mining
The robustness checks discussed above do not account for the possibility that
our results may be attributed to data mining. Our study is largely motivated by
the political cycles literature in macroeconomics,17 which prompted us to test the
hypothesis that political cycles have an impact on stock market returns. Inspired
17
See Alesina and Rosenthal (1995), Alesina et al. (1997), and references therein.Political Cycles and the Stock Market 1857
Excess Value Weighted Returns Excess Equally Weighted Returns
60 60
40 40
Annualized Percentage Difference in Returns
Annualized Percentage Difference in Returns
20 20 95% CI
95% CI
0 0
Mean
Quantile Quantile
Mean
-20 -20
-40 95% CI -40
95% CI
-60 -60
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
τ−Quantile of Difference in Returns τ−Quantile of Difference in Returns
Figure 2. Quantiles of di¡erence in returns under Republican and Democratic
presidencies, 1927 to 1998. Figure 2 displays the di¡erence in quantiles of value-weighted
and equal-weighted returns between Republican and Democratic presidents.The quantile
of the di¡erence is computed as
X
^ ðtÞ ¼ arg min
½^aðtÞ; b rt ðrt a bRDt Þ
a;b
t
as discussed in the text (equation (4)). The di¡erence in quantiles b ^ðtÞ is displayed as a
solid line, for t ¼ 0.02,y,0.98, whereas the dashed pattern denotes the 2.5 percent, mean,
and 97.5 percent of the estimates, computed by bootstrap. The mean of the di¡erence is
shown for reference as a light straight line. The signi¢cance of the di¡erence comes from
the middle quantiles, which supports our robustness claims.
by that literature, rather than by a clearly formulated theoretical model, we tried
to correlate political variables with stock market returns. In addition to the pre-
sidential cycle variable studied in the paper, we also tried other variables related
to the party in control of Congress. Since those results were largely insigni¢cant,
in the interest of brevity we do not report them in the paper. However, we need to
acknowledge that the results that we do present were preceded by this modest
search for a statistically signi¢cant variable. We should therefore adjust the dis-
tribution of test statistics to take this search into account.1858 The Journal of Finance
Based on the work of Bonferroni, several authors18 propose to adjust the
p-values of statistical tests by multiplying them by the number of hypothesis
tested.19 Considering that we also looked at Congress besides the presidency, we
should at least double the p-values of our tests. Making this adjustment in Table II,
Panel B, for example, the di¡erence in value-weighted excess returns between par-
ties would have p-values of 0.08, 0.18, and 0.10 in the full sample and the two sub-
samples, respectively. We therefore conclude that data mining is indeed a concern
in this study.
Unfortunately, there is a serious drawback to the Bonferroni approach. Although
the chance of incorrectly ¢nding an e¡ect (or making a type I error) on an indivi-
dual test is reduced, the chance that no e¡ect is found, while in fact there is an
e¡ect (or making a type II error) is increased. Thus, the size correction comes at
the cost of power loss. As we point out in Section II, power is already low in our
tests due to the low number of presidencies. If we add the Bonferroni correction,
the ability to ¢nd any variable (political or other) that forecasts the stock market is
virtually eliminated.We therefore choose to present the test statistics without any
adjustment for potential data mining problems and let the forewarned reader de-
cide on the signi¢cance of our reported p-values. Ultimately, the concern of data
mining can only be dispelled after we accumulate enough out-of-sample data.
IV. Expected or Unexpected Returns?
Having established that there is indeed a di¡erence in returns between Repub-
lican and Democratic presidencies, we proceed in this section to investigate
whether this di¡erence in realized returns can be attributed to a di¡erence in (ex
ante) expected returns or to a di¡erence in unexpected returns. A di¡erence in
expected returns would be consistent with a higher risk premium charged by the
market for Democratic presidencies. In contrast, if the di¡erence is due to unex-
pected returns being higher under Democrats, that would signal that the market
is systematically positively surprised by the policies of Democratic presidencies.
A. A‘‘Proxy’’ Explanation
The most natural explanation for the correlation between presidential-parti-
san variables and excess returns is based on a ‘‘proxy’’e¡ect. The presidential cy-
cle might merely be proxying for variations in expected returns due to business
cycle £uctuations. Since variations in returns have been associated with business
cycle £uctuations,20 and business cycle £uctuations have been associated with
18
See Leamer (1978) and references therein.
19
A bootstrap-based procedure for dealing with data mining was recently proposed by Sul-
livan et al. (2001). This approach is di⁄cult to implement in our case, since it requires that we
¢rst specify all possible political variables and investigate whether they correlate with the
stock market. Unfortunately, the universe of variables related to politics is impossible to
enumerate.
20
See Campbell (1991), Fama (1991), and Campbell, Lo, and MacKinlay (1997) for a textbook
treatment.You can also read
